DeepReason CLAUDE.md

Repository instructions for DeepReason, a reasoning tool that records model conjectures, criticism, decisions, and capability changes in files that can be replayed and verified.

In plain words
What is it for?
Use them when diagnosing problems, planning changes, implementing fixes, or verifying results in DeepReason.
Why use it?
They help agents choose the right workflow and preserve an auditable record instead of treating model-written explanations as proof.

Instructions file

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add instructions/ahepi/deepreason/claude-md
Clone the repo
git clone --depth 1 https://github.com/AHepi/DeepReason
Per session 8,902 This file is loaded in full into every session.
When invoked 8,902 The same file — it is already loaded in full.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.08902 $0.08902
Opus 5 $0.04451 $0.04451
Sonnet 5 $0.01780 $0.01780
Haiku 4.5 $0.00890 $0.00890

Measured 2d ago against content hash 73c4ebf494f3, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

DeepReason CLAUDE.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

CLAUDE.md · 636 lines

How it starts

The opening of the file, as written. The whole thing — 636 lines — stays where its author put it; the contents beside it link to each section on GitHub.

CLAUDE.md — operating DeepReason

DeepReason is a Popperian reasoning harness: it drives a provider model (currently glm-5.2 on Ollama Cloud) through conjecture–criticism cycles over an append-only, replay-verifiable record. Everything meaningful is TYPED — stops, denials, refusals, capability lifecycles — and the record is the only admissible evidence about what a run did. Model prose is never evidence; log.jsonl, objects/, progress.jsonl, run-status.json, REPLAY_VALIDATION.json, and verify_root are.

Which workflow to use

Both families now begin with a MAP PREFLIGHT: resolve the work to DR-SUB-/DR-CON-/DR-SEAM- ids from docs/map/INDEX.md, read the seam before the subsystems, and read INV-frozen-surfaces.md before designing. Record the ids in the tranche's first artifact so every later phase starts from the same map.

Two skill families live in .claude/skills/. Route ALL substantive work through one of them — they exist to prevent scope creep, missed steps, and forgotten inputs:

  • Something is broken / suspiciousdeepreason-orchestrator (phases: dr-set-goal → dr-diagnose → dr-reproduce → dr-propose-fix → dr-implement-fix → dr-verify-outcome). Diagnosis comes from the typed record BEFORE code reading.

  • The operator suggests a changedr-change-orchestrator (phases: dr-capture-request → dr-spec-change → dr-plan-steps → dr-execute-step (one step per invocation) → dr-validate-change → dr-deliver-change). Authority is the operator's verbatim words, ledgered in REQUEST.md; every artifact traces to requirement numbers.

  • The operator asks what is broken / unused / out of datedr-audit-orchestrator (dimensions: broken, dead, docs-drift, spec-drift, goal-trace). Read-only: produces AUDIT_REPORT.md plus a ready-to-send fix prompt per finding; every verdict compares against docs/AUDIT_BASELINES.md. Rated for inexpensive models — every step is a command, a paste, or a baseline comparison.

Cross-routing: a defect found mid-change is PARKED, not fixed; a change wished for mid-defect is PARKED, not implemented. One tranche, one goal. The audit family never fixes anything anywhere — findings become parked prompts for the other two families.

Read the full file on GitHub · 636 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 636 lines · 8,902 tokens per session scan A 73c4ebf494f3

Subscribe to this mod's changes

DeepReason CLAUDE.md is an instructions file published in the GitHub repository AHepi/DeepReason (140 stars, last pushed 2d ago), licensed MIT. It adds 8,902 tokens to every session, about $0.0445 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.